Nature Neuroscience

Functional magnetic resonance imaging (fMRI) contains rich individual, cognitive, and pathological information, yet no universal model exists for multi-task modeling of these dimensions. Here, we introduce MAMBAxBrain, a multi-task neural framework that integrates Mamba architecture with functional connectivity analysis to jointly model the temporal dynamics and spatial coordination of neural activity. MAMBAxBrain achieves high accuracy across four distinct fMRI objectives--brain fingerprinting, cognitive task decoding, reaction time prediction, and schizophrenia classification--consistently outperforming state-of-the-art methods with robust crosssession generalization. Interpretability analyses show that each task engages distinct, biologically plausible circuitry--from higher-order association cortex for identity to subcortical-motor loops for reaction time and disrupted control-sensory connectivity for schizophrenia. These findings inform a longstanding debate: rather than operating through wholly separate or entirely shared systems, the brain preferentially recruits task-specific circuits while retaining common representational structure across functions.

Challenges in verbal communication are a prominent feature of autism. However, gene regulatory programs in speech-related cortical regions remain poorly characterized. In parallel, it remains unclear whether the heterogeneous genetic factors underlying autism converge on shared neurobiological mechanisms. To address these gaps, we generated paired transcriptomic and epigenomic data from post-mortem human brain tissue across 100 donors. Here, we show that transcriptional differences in the speech-related Brodmann Area 22 in individuals with neurodevelopmental conditions, including autism, are strongest among those with a known genetic diagnosis. A similar but attenuated signature is observed in those without a genetic diagnosis. These transcriptional differences are most pronounced in neurons, with glutamatergic L4/5 intratelencephalic neurons affected across multiple modalities. Finally, multimodal analysis implicates altered RFX3-dependent networks as a central hub in autism, particularly among L4/5 intratelencephalic neurons in non-verbal individuals. Together, our study identifies regulatory architecture linking chromatin state, transcriptional output, and variation in verbal ability in autism.

1.Schizophrenia spectrum disorders (SSDs) present a profound clinical enigma, manifesting as a heterogeneous continuum ranging from the chaotic volatility of acute psychosis to the impenetrable rigidity of systematized delusions. While neurobiological research has independently implicated NMDA receptor hypofunction or dopaminergic hyperfunction as cardinal pathophysiological distinct mechanisms, a computational framework capable of bridging these distinct cellular deficits to the spectrums vast phenomenological diversity remains elusive. Here, we propose a biologically plausible neural model using a dynamic Bayesian inference with separable positive and negative prediction-error pathways. We demonstrate that NMDA hypofunction selectively blunts negative prediction errors, fostering rigid, bias-dominated beliefs, while dopaminergic hyperfunction uniformly amplifies error signals, driving volatile, observation-dominated states. Their interaction reconstructs SSDs as a continuous bias-volatility spectrum, accounting for key neurophysiological markers and offering a theoretical foundation for mechanism-based patient stratification.

Flexible behavior requires binding actions to their underlying cognitive variables, yet classical basal ganglia models emphasize a serial architecture where striatal action selection precedes a pallidothalamic motor gate. We tested this framework by recording single-neuron activity across the human pallidothalamic circuit while participants reported perceptual decisions with confidence-weighted reaching movements. Neurons in globus pallidus externus and internus, subthalamic nucleus, and motor thalamus exhibited sharp activity increases at movement onset that persisted through execution and feedback. Choice and confidence were continuously encoded before, during, and after movement. Population decoding revealed a transformation along the circuit, from dynamic upstream representations to stable, low-dimensional representations downstream, particularly in motor thalamus. After feedback, outcome signals emerged selectively in subthalamic nucleus and motor thalamus. In motor thalamus, confidence and outcome converged into a representation consistent with unsigned reward prediction error. These findings redefine the human pallidothalamic circuit as an interface that binds decision, action, and evaluation during behavior.

Axons are the brains wiring, organized into bundles that connect nearby and distant regions. Axon caliber determines signal conduction velocity and varies both within and across bundles, reflecting the brains diverse functional demands. Much of what we know about this organization derives from 2D histology, assuming cylindrical axons whose calibers are described by their radius. Yet, recent 3D histology reveals that the radius varies along an individual axon--with implications for both characterizing axon caliber and potentially conduction velocity predictions. We show in 450,000 3D rat axon reconstructions that--despite this individual variation--axon bundles possess stable radius distributions at the ensemble level, which 2D cross-sections faithfully represent. This representativeness extends to conduction velocity predictions, as along-axon variation has only modest impact. In particular, large axons exhibit especially stable conduction, emphasizing their key role in time-critical signaling. With 2D sampling validated, we leverage 46 million human corpus callosum axons from 2D histology to determine sample size requirements across neuroscience applications. Our findings reinforce decades of 2D histology-based research on axon organization and its functional implications, while guiding future study design.

Across many brain areas, neurons produce heterogeneous, seemingly disordered responses. Yet the circuits these neurons comprise cannot be purely random; they must possess some structure to generate representations and computations underlying behavior. How much structure is present in recurrent connectivity relative to disorder, and how the interaction between them shapes population dynamics and single-neuron responses, remain incompletely understood. Recurrent neural networks trained to perform tasks have become a leading model of such circuits, but conventional training yields a single point in a vast space of task-compatible solutions, with no systematic way to explore this space and no theory of how internal representations vary within it. Without such a theory, the questions above cannot be addressed, and comparisons between trained networks and neural data are difficult to interpret. Here, we introduce a control parameter governing the degree to which learning reshapes recurrent connectivity and permits interpolation between a reservoir regime and one in which recurrent weights are restructured by learning to produce task-relevant internal representations. Varying this parameter generates a family of task-compatible solutions whose internal dynamics differ in a controlled and interpretable way. We derive a dynamical mean-field theory showing how the balance of randomness and structure in connectivity influences both population-level dynamics and single-neuron responses, with the former converging to a deterministic limit and the latter becoming independent samples from a singleneuron response distribution. When connectivity is random, this response distribution is Gaussian; recurrent restructuring drives it toward task-dependent, non-Gaussian forms. In linear networks, restructuring amplifies task-relevant frequencies; in nonlinear networks, it drives a phase transition from chaotic, high-dimensional activity to ordered, low-dimensional dynamics that generalize temporally beyond the training period. We apply this theory to a reaching task in which a recurrent network must reproduce macaque muscle activity. Optimally matching simultaneous motor-cortex recordings requires a relatively small degree of restructuring in which learned structure coexists with random heterogeneity. These results are suggestive of a broader picture in which large recurrent neural circuits are quite random, but contain, to varying degrees, structured recurrent connectivity sufficient for generalizable, task-relevant representations1.

A defining feature of substance use disorder is that repeated drug use does not always lead to addiction, motivating the search for biomarkers of vulnerability1. Reduced striatal dopamine D2/3 receptor availability is a robust PET correlate of problematic stimulant use2-5, but the signal may reflect high endogenous dopamine level, and it conflates presynaptic D2 autoreceptors on dopamine axons with postsynaptic D2/3 heteroreceptors on striatal projection neurons. We dissociated these contributions using cell type-specific Drd2 haploinsufficiency in dopamine neurons (autoD2KD), D2-expressing medium spiny neurons (MSN-D2KD), or both. Autoreceptor haploinsufficiency (autoD2KD) weakened presynaptic control of dopamine release, enhanced phasic gain, and prolonged cocaine-evoked dopamine elevations. This was accompanied by a hyper-exploratory trait and altered cocaine adaptation. Specifically, autoD2KD mice showed greater cocaine-seeking behavior, despite intact responses to sucrose reward and punishment. Although all genotypes showed graded reductions in striatal D2/3 binding, D1-like compensations diverged, resulting in different D1:D2/3 ratio in the striatum. The clinical implication is that striatal D1 density and D1:D2/3 balance may emerge as critical biomarkers for distinguishing cell-type-specific D2 reductions relevant to addiction vulnerability.

Face-to-face communication relies on the seamless integration of visual and acoustic cues, yet the spatiotemporal principles governing how the human brain dynamically represents and combines these multisensory streams remain largely unresolved. To address this, we recorded high-density electrocorticography (ECoG) from eight participants perceiving matched audiovisual, audio-only, and video-only continuous natural Mandarin speech. Using time-frequency-resolved encoding models, we reveal complementary, frequency-dependent integration regimes across the temporal lobe. We show that the superior temporal gyrus (STG) implements a feature-selective, auditory-dominant strategy, utilizing visual input to selectively strengthen low-frequency representations of lip-reading kinematics. Conversely, the middle temporal gyrus (MTG) acts as a higher-order multisensory hub, employing a frequency-selective strategy to broadly integrate diverse facial and articulatory features. Crucially, we demonstrate that access to visual information during perception significantly improves the acoustic and lexical accuracy of neural speech decoding and re-synthesis, with the MTG driving the largest gains in linguistic intelligibility. These findings uncover the dissociable neural architectures supporting robust multisensory perception, providing critical mechanistic insights for the development of next-generation, multimodal brain-computer interfaces.

Sleep is thought to stabilize newly formed memories, yet the neural reorganization through which sleep converts learning-induced plasticity burden into stable memory remains unclear. Current state-specific oscillatory markers provide limited insight into how learning reshapes population dynamics across sleep-wake states. Using spectral parameterization of high-density EEG, we show that declarative learning redistributes frontocentral waking aperiodic regimes toward flatter slopes relative to a non-learning control. These deviations are renormalized during subsequent NREM sleep toward steeper slopes with accompanying oscillatory power shifts. Spatial deviations in waking slopes reveal a region-specific coupling with NREM, dissociable from canonical oscillatory signatures. A latent neural-memory mode showed that the wake-sleep aperiodic contrast best predicted overnight accuracy changes, whereas local oscillations and aperiodic shifts defined the spatial pattern of neural variation supporting memory stabilization. Together, these findings identify sleep-dependent recalibration of learning-perturbed population dynamics as a systems-level mechanism linking homeostatic plasticity to memory consolidation.

Experience unfolds as a stream shaped by hidden causes that change over time. Adaptive behavior requires inferring the underlying states and adjusting when they change. Yet, how neural circuits discover and track latent states remains unclear. Here we introduce NeuraGEM, a neural architecture that combines fast transient activity with slow synaptic plasticity to implement an online analogue of Expectation-Maximization. By separating timescales, NeuraGEM clusters sequential experiences, detects context changes, and stabilizes task-specific computations. The model generalizes beyond conventional recurrent networks and reproduces key features of human contextual learning, including curriculum-dependent effects. It also gives rise to population dynamics resembling those observed in brain circuits, including line-attractor structure and transient error responses at change points. Together, these findings provide a mechanistic account of how neural circuits organize experience into latent states that support rapid inference and adaptive behavior.

The embedding of low-dimensional latent states in the activity of large neuron populations has become a tenet of systems neuroscience. Despite the stability of these latent representations over time,1 the underlying activity of individual neurons is known to change both within experimental sessions and across days;2 yet, less attention has been devoted to changes in the coordinated activity of neuron populations on short timescales, particularly under conditions where networks must adapt quickly (e.g., during learning or after injury). To investigate, patterns of individual neuron contributions to population state dimensions in motor cortex were tracked over short blocks of repeated reaching trials. The number of distinct encoding patterns was consistently less than the typical dimensionality reported for motor cortex. Although the neuron population state space and dynamics were effectively conserved, these underlying encoding patterns changed in two ways: fast switches among themselves, as well as slow modifications over time. To explore whether these two drift timescales shared a common physiological mechanism, we analyzed the response of motor cortex to a causal perturbation. Direct electrical current was passed through two recording electrodes to terminate a small number of neurons,3-6 which evoked drift over both fast and slow timescales, both together and independently. While changes in fast switch rates were not necessarily associated with behavioral deficits, a significant increase in slow drift was accompanied by a decrease in behavioral performance. Together, these results reveal an additional timescale of drift in correlated population activity that could help guide the discovery of cellular and network mechanisms responsible for the maintenance, (re)learning, and recovery of low-dimensional structure in neuron populations.

Neuronal ensembles--groups of neurons that exhibit coordinated activity during behavior--are a fundamental feature of cortical computation. Dendritic branches amplify clustered synaptic inputs through local nonlinearities, suggesting that presynaptic groups might organize their connections in specific spatial patterns to engage these mechanisms. Whether the same axon groups form synaptic clusters with consistent spatial arrangements across different target neurons remains unknown, but nanoscale connectomes would resolve such anatomical motifs if they exist. We analyzed millions of synaptic connections in a connectome of mouse visual cortex and found over 700,000 axon groups that repeatedly cluster their synapses onto dendritic branches of multiple pyramidal cells, with over 500,000 maintaining consistent distal-to-proximal arrangements. These repeated patterns occur far more frequently than expected from spatial proximity or layer-based connectivity rules. Axon groups preferentially target specific dendritic branches and position their synapses in stereotyped spatial configurations across multiple postsynaptic partners, revealing that functional ensembles leave characteristic anatomical signatures in cortical microarchitecture.

Among individuals with equivalent Alzheimers pathology, cognitive outcomes can diverge by decades, a phenomenon termed cognitive reserve that remains descriptive after thirty years of research. We propose that the [~]109-to-10 bits/s gap between sensory input and behavioral output functions as error-correcting redundancy in the sense of Shannons channel coding theorem. Progressive neuronal loss maps to symbol erasure in a redundant code, and the critical damage fraction at which cognition fails is dc = 1 - k/n, where k {approx} 10 bits/s is the behavioral channel requirement and n is the effective number of coding units. We evaluate this threshold across three channel models (binary erasure, Gaussian, and Erd[o]s-Renyi percolation) and show that all produce a sharp phase transition from reliable to unreliable decoding. The framework makes four testable predictions: (i) dc scales with the measurable redundancy ratio{rho} = n/k, which accounts for clinical heterogeneity; (ii) information-theoretic redundancy from resting-state fMRI should predict time-to-conversion beyond structural atrophy; (iii) the decline trajectory near dc is sharp, consistent with the "cognitive cliff"; and (iv) motor circuits, operating at higher bandwidth, have lower reserve than cognitive circuits. Significance StatementCognitive reserve (why some brains resist dementia pathology better than others) has been described for thirty years but never given a quantitative, information-theoretic foundation. We propose that the roughly hundred-million-fold gap between sensory input ([~]109 bits/s) and behavioral output ([~]10 bits/s) functions as error-correcting redundancy in the Shannon coding-theoretic sense. This yields a closed-form critical damage threshold, dc = 1 - k/n, below which cognitive function is preserved and above which it collapses; this is consistent with the clinically observed plateau-then-cliff pattern of dementia. The framework unifies cognitive reserve with channel coding theory, accounts for individual heterogeneity in disease onset, and generates falsifiable predictions that link information-theoretic redundancy measures to time-to-clinical-conversion.

Stroke is one of the leading causes of global disability, yet the principles governing how focal brain injury disrupts large-scale neural connectivity over time remain poorly understood. Here, we leverage a longitudinal multimodal dataset to track the evolution of individual-specific connectivity patterns, or brain fingerprints, over the first year after stroke. Despite a persistent shift from healthy architecture, we demonstrate that each patients unique functional connectome fingerprint is remarkably resilient and stabilizes within three weeks. This early global stabilization masks a protracted system-specific reorganization of brain circuits, which is characterized by an initial increase in connectivity within sensory and attention systems, followed by a decline across higher-level association networks. A joint structure-function embedding further shows that recovery involves a gradual shift toward the normative healthy range, driven primarily by functional reconfiguration atop a stable structural lesion. Crucially, a multivariate prediction model reveals that early functional signatures selectively forecast long-term impairment in language, executive function, and attention. Together, our results define the post-stroke brain as a shifting but constrained dynamical system, identifying early-stabilized brain patterns as biomarkers for individual recovery profiles and targets for personalized neurorehabilitation.

Models of cognition often treat fast electrical dynamics as separate from slower associative plasticity. We asked whether rapid synaptic structural plasticity is required for coordinated cortical processing during wakefulness. We developed SynC, a chemogenetic perturbation that, upon A/C administration, reversibly blocks associative dendritic spine enlargement and impairs motor learning without detectably altering baseline synaptic/neuronal physiology. SynC activation in frontoparietal cortex reduced laser-dot chasing and delayed feeding initiation, and reduced wake-state stability. Population recordings revealed a locomotor-active but impaired regime with preserved mean activity and {gamma} power but reduced pairwise correlations during quiet, immobile epochs, with increased network dimensionality (Interm-C), intermittently interrupted by an arrest state (State-C) with suppressed {gamma} power. Thus, intact Rac1-dependent spine-enlargement mechanisms are acutely required to maintain the functional coupling necessary for coordinated cortical processing during wakefulness. One sentence summaryAcute, reversible in vivo blockade of associative spine enlargement shows that rapid structural plasticity is required to sustain cortical processing in EEG/EMG-defined waking.

Linking synaptic-level perturbations to distributed brain-network dynamics remains a central challenge for understanding and treating mental illness. Although recent whole-brain models can reproduce individual brain activity patterns, they largely function as descriptive simulators rather than mechanistic, intervention-capable systems. Here we present an intervention-capable digital twin of the human brain, integrating individual neuroanatomy and task-evoked dynamics within a neuronal-scale framework. Individualised digital twin brains recapitulate a participant-specific compact cortico-subcortical network phenotype that captures transdiagnostic psychopathology across population and clinical cohorts. In silico modulation of excitatory and inhibitory synaptic conductance produces bidirectional, heterogeneous network responses across individuals. Population-scale simulations stratify individuals and predict longitudinal symptom trajectories from DTB-derived response profiles. Independent pharmacological functional MRI data further validate the predicted baseline-dependent network responses in vivo. Together, these findings establish digital brain models as experimental platforms for mechanistic perturbation, behavioural prediction and stratification, providing a foundation for precision neuroscience and psychiatry.

Systems-level consolidation holds that the hippocampus rapidly encodes new information during wakefulness, and that coordinated cortico-hippocampal replay during subsequent sleep transfers and stabilizes and refines those traces in cortex. This idea captures key learning principles, but exactly how replay reshapes the synaptic-weight landscape - creating new representations while preserving old ones - remains unclear. To address this, we used a biophysically realistic network model to probe the effects of slow-wave sleep (SWS) on synaptic-weight space. We show that previously learned memories are stable attractors in that space, and that hippocampus-driven interactions between sharp-wave ripples and cortical slow waves push the system into new attractor states that jointly encode old and new memories. As a result, replay allows recently acquired information to be incorporated without degrading prior memories. Our results offer a novel mechanistic - and conveniently "geometric" - framework for understanding how sleep-driven replay sculpts synaptic weights during consolidation. SIGNIFICANCE STATEMENTStoring, processing, and retrieving information underpins intelligent behavior. Sleep extracts invariant features from prior experience, promoting the emergence of explicit knowledge and insight. Yet despite abundant empirical findings, our understanding of how sleep reshapes memory representations across brain networks remains limited. Here we present a novel framework that describes how memories are encoded in synaptic-weight space and how sleep dynamics reorganize that landscape. These results advance our understanding of how the brain solves core problems of lifelong learning.

Large-scale electrical perturbation of the human brain provides a unique model for understanding how multiscale biological constraints shape behaviorally relevant reorganization. Here, we integrate longitudinal neuroimaging coordinates from 148 experiments ({approx}2,300 subjects) with normative connectomics, chemoarchitecture, intrinsic electrophysiology, and transcriptomics to identify cross-scale principles governing human brain reconfiguration under strong perturbation. Convergent hubs of structural and functional plasticity embed within default-mode and salience systems and show complementary coupling to visual networks, linking perturbation-induced change to large-scale circuits supporting affective regulation, memory, interoception, and psychosis-relevant processes. These macroscopic patterns align with intrinsic cortical dynamics and chemoarchitectural gradients dominated by 5-HT1A receptors, with additional contributions from D2, -opioid and GABAA systems, and are enriched for astrocytic and microglial gene expression, implicating glial plasticity in systems-level reorganization. Finally, in a separate intervention dataset, regularized statistical-learning models demonstrate that this multiscale signature tracks behaviorally relevant symptom change specifically under strong electrical perturbation. Together, these results outline general organizing principles linking molecular, cellular and network-level constraints to human behavioral adaptation, providing a computational framework for understanding how large-scale perturbations reshape brain systems across levels of biological organization.

Spontaneous activity in the human cortex is organized along large-scale functional gradients, yet how these macroscale patterns relate to laminar functional architecture remains unclear. Here, we use sub-millimeter, layer-resolved 7T resting-state fMRI to map whole-brain functional connectivity gradients across cortical depth. We represent the cortex as a multilayer network with distinct connectivity profiles for deep, middle, and superficial layers, and derive laminar dissimilarity indices that quantify differences between layer-wise gradient embeddings within and across regions. These indices systematically distinguish canonical resting-state networks and vary along a histology-informed microstructural axis that captures a canonical sensory-to-limbic gradient of cytoarchitectural differentiation. Layer-specific analyses further show that the balance between deep- and superficial-layer dissimilarity tracks a functional hierarchy estimated from independent effective-connectivity modelling, consistent with feedback-like intrinsic connectivity at rest. Together, the results establish laminar gradient-derived indices as a bridge between cortical microstructure, large-scale network organization, and hierarchical information flow in the human cortex.

Neuromodulatory systems such as dopamine and acetylcholine enable cortical circuits to rapidly shift between behavioral states, yet it remains unclear how receptor-specific signaling reshapes what single neurons compute. Here we combined whole-cell recordings from excitatory and inhibitory neurons in layer 2/3 of mouse somatosensory cortex with information-theoretic analyses under frozen-noise stimulation. We quantified stimulus-to-spike information transfer and profiled each neuron across four functional domains: passive membrane biophysics, action potential dynamics, intrinsic adaptation currents, and input feature selectivity captured by spike-triggered averages. Activation of D1, D2, or M1 receptors produced cell-type and receptor-specific changes in fractional information and firing output, demonstrating that neuromodulators regulate encoding beyond simple gain control. Unsupervised clustering revealed that receptor activation reshapes neuronal functional identities, consistent with a dynamic computational phenotype. To identify the organizing principle underlying these transitions, we applied multi-set correlation and factor analysis and found that neuromodulation systematically reorganizes covariance among functional domains. In excitatory neurons, dopaminergic activation led to a decreased coordination between input feature selectivity and other functional properties while strengthening coordination among response based features such as spike dynamics and adaptation. Inhibitory neurons, by contrast, generally exhibited increased coordination across domains. These findings identify neuromodulation as a reconfiguration signal that reshapes not only individual cellular properties but also the architecture linking them, thereby dynamically expanding the computational repertoire of cortical circuits.